Machine-learned interatomic potentials (MLIPs) are becoming increasingly important in computational chemistry. These tools enable predictive simulations of the behaviour of matter on the scale of atoms and bonds. Building an MLIP model for a new material has, however, traditionally been a complex and time-consuming process, often requiring expert oversight.
In a new study published today in Nature Communications, researchers from Oxford Chemistry and the Federal Institute for Materials Research and Testing (BAM) in Berlin present autoplex, an open-source software package which helps researchers to quickly build MLIPs for a wide range of systems in a largely automated fashion. The MLIPs can then be used to run atomistic simulations of the material’s behaviour.
The team developed flexible and modular workflows that can create robust MLIPs from scratch within a matter of days. Their approach uses iterative random structure searching: starting from just randomised atomic configurations, initial MLIP models are fitted and used to search for lower-energy structures (for example, crystalline polymorphs of a material). This idea was pioneered in 2018 by Deringer et al., and has now been implemented in workflows that leverage the atomate2 computational ecosystem. This means that the approach can now be deployed efficiently on supercomputers, which reduces the amount of manual input and effort required from users of the method.
The new paper includes many examples that showcase the method’s applicability to a range of systems with complex structures, including: crystalline modifications of silicon and titanium oxides, from the most fundamental solid-state structures to more complex polymorphs; liquid water; and phase-change memory materials used in digital data storage.
Dr Yuanbin Liu, a postdoctoral researcher at Oxford Chemistry and first author of the work, commented:
Building MLIPs from scratch has traditionally been a major hurdle. With autoplex, our new computational ecosystem, we hope to streamline the entire process by minimising both the required domain knowledge and the amount of manual effort involved. Our primary goals in designing autoplex were to make it user-friendly and easily extensible to adapt to diverse research scenarios. We’ve had great success testing it on a variety of materials and are excited to see how the community will build on and benefit from it.
The project is part of a wider collaboration between two groups with complementary strengths: atomistic machine learning in the Deringer group in the Inorganic Chemistry Laboratory at Oxford, and automation and workflows for materials science at BAM, led by Prof Janine George. The autoplex code and its documentation are openly available at https://github.com/autoatml/autoplex.
Prof Volker Deringer, senior author of the study, commented:
MLIPs are on their way to becoming mainstream simulation tools in chemistry, and efficient, robust, and reproducible workflows are set to be a key part in that. I’m delighted to see our collaboration play a part in this important wider effort.
The research at Oxford was generously supported through a UKRI Frontier Research grant.
The paper is openly available at https://doi.org/10.1038/s41467-025-62510-6.
